The Illusion of Intervention: Your LLM-Simulated Experiment is an Observational Study

This paper shows that LLM-based “synthetic user” experiments often do not implement randomized trials; they behave like observational studies with intervention-dependent selection bias. The authors formalize “user drift” as a causal confounding problem: when a persona L is held fixed, the latent user state X \ L can still shift differently under treatment A = 0 versus A = 1, so the observed effect τ_obs estimates a mixture of treatment effect and selection bias rather than the average treatment effect. They diagnose drift with negative control outcomes Z—attributes that should be invariant to the intervention—and quantify distributional shift with total variation distance between P(Z | A = 1, L) and P(Z | A = 0, L). Across six models, including Qwen3-30B-A3B, GPT-OSS-20B, Gemma-4-31B-it, Gemma-3-4B-it, and Gemini 3 Flash, they find substantial baseline TVD in OpinionQA, Book Opinions, and MovieLens, with the smallest Gemma-3-4B showing especially large drift. They then mitigate bias by iteratively eliciting additional confounders L′, especially task-relevant variables such as political ideology, media habits, or book/movie selection behavior. This adjustment often reduces TVD and stabilizes effect estimates, but generic demographics can initially worsen bias before targeted confounders bring it down. The key novelty is methodological: the paper argues that persona prompting plus intervention is not enough for causal identification, and that internal validity in LLM simulations requires negative controls and confounder adjustment, not just better mimicry of human responses.